AIxCrypto launched RoboShare, a robot rental marketplace, alongside a new infrastructure platform for autonomous assets. The company introduced these tools at the Automate Show in Chicago. This initiative aims to extend the commercial utility of robots beyond their initial purchase. RoboShare facilitates the renting of various robotic units. Furthermore, it establishes an autonomous asset network to support diverse AI agents and machinery.
Develops an end-to-end approach for creating policies for humanoid robots using NVIDIA Isaac GR00T. Addresses the growing need for repeatable workflows as teams advance beyond initial robot build-out. The platform facilitates task-specific skill development for these complex machines. This capability streamlines the process of programming and testing advanced robotic behaviors.
RoboCup took place in Incheon, South Korea, from July 2nd to July 6th. Participants engaged in various activities including competitions, training sessions, and a symposium. The social media round-up documented the actions of the event's participants. Furthermore, the content announced the official start date for RoboCup 2026.
IEEE recognized robotic pioneer Toshio Fukuda with its Richard M. Emberson Award for advancing robotics. His extensive contributions include developing biomedical and industrial robotic systems, micro-nano robotics, and mechatronics. Fukuda pioneered modular, cellular robotic systems (CEBOTs), allowing them to assemble complex structures like interlocking bricks. He also helped develop brachiation robots mimicking monkey movement for inspecting infrastructure. A highly prolific scholar, he served as IEEE president in 2020 and is currently affiliated with the Egypt-Japan University of Science and Technology.
NVIDIA has integrated its robotics tools with Hugging Face's platform through a standardized pipeline called LeRobot. This integration incorporates several key components from NVIDIA's suite. Specifically included are NVIDIA Isaac GR00T 1.7 and Isaac Teleop functionality. Furthermore, the upcoming Cosmos 3 world model will also be made available within this ecosystem. This collaboration standardizes open-source robotics development using these powerful tools on Hugging Face.
Modern manufacturing facilities comprise vast, connected spaces requiring comprehensive oversight. Engineering teams struggle with maintaining clear visibility across diverse operational zones, including factory floors, plant rooms, and external assets. The increasing complexity of these sprawling environments demands better monitoring capabilities for continuous equipment operation. Robot dogs are being highlighted as a leading solution to address this need for broad site awareness. These robotic units provide crucial real-time intelligence across large physical footprints within industrial settings.
ABB Robotics has expanded its Autonomous Mobile Robotics (AMR) line with the launch of the Flexley Stack F712. This new product establishes a fully interoperable ecosystem across all primary Visual SLAM AMR types offered by the company. The platform integrates autonomous forklifts, tugs, and movers into one unified system. This comprehensive capability allows customers to automate a wider variety of material-handling and intralogistics processes.
Automate 2026, North America's largest robotics and automation event, achieved record attendance. Over 50,000 registrants and 1,230 exhibitors participated at McCormick Place in Chicago. The association for Advancing Automation (A3) hosted the successful event from June 22-25. The show covered cutting-edge technologies across robotics and automation sectors. It utilized a substantial 425,000 square feet of exhibition space.
ExRobotics introduced its UL Certified ExR-2.5 autonomous inspection robot in North America. The launch occurred at the Energy Drone & Robotics Summit in Houston. Achieving this certification represents an industry first for the company. The ExR-2.5 is unique as the only UL Certified inspection robot designed for potentially explosive industrial environments. This provides operators with a recognized method for site inspections.
HelloFresh has significantly expanded its refrigerated fulfillment capabilities by incorporating Locus Robotics' automated cold-storage hardware. This modification increases their chilled SKU capacity from 100 to 500. The fivefold increase supports both greater meal variety offerings. Furthermore, this expansion unlocks new revenue opportunities across the entire HelloFresh brand portfolio. One of the brands benefiting is Factor.
RLWRLD, a physical AI company, launched "All Hands Up!" to benchmark dexterous robotic hands. This open web platform offers technical reports and visualization tools based on real-world operation data. The foundation for this work is RLWRLD's proprietary robotics model, RLDX-1. The platform aims to allow rigorous analysis of various commercially available dexterous robot hands.
Flytrex successfully scaled its shared airspace operations, managing thousands of coordinated drone flights monthly. This milestone demonstrates safe operation within complex aerial environments. The company achieved this feat using advanced technology for managed air traffic. It solidifies Flytrex's position as a leading commercial drone operator in the United States. No reported airspace conflicts were experienced during these high-volume operations.
Robust.AI has chosen Aptiv's intelligent perception solutions for its next-generation Carter warehouse robot. These selected solutions include AI and Machine Learning based sensor fusion, which is powered by the Aptiv Pulse sensor. This integration will enhance the capabilities of Robust.AI's collaborative mobile robot. The decision follows an existing partnership between the two companies. The incorporation of Aptiv's proven technology marks a significant step in the robot's development.
Lumos Robotics' Prime R0 industrial embodied AI model secured the top spot on the latest MolmoSpaces leaderboard. The Chinese company reported that its 2.8-billion-parameter model outperformed larger competitor models, including those from Nvidia. Its high performance was demonstrated across both single-arm fine manipulation and dual-arm collaboration tasks. Achieving this ranking solidifies Lumos Robotics' state-of-the-art capabilities in the field.
South Korean robotic hand developer Tesollo has initiated preparations for an initial public offering (IPO). The IPO will be through Korea's technology-special listing track, with a planned listing following 2027. In March, the company appointed KB Securities to serve as its lead underwriter. Furthermore, Tesollo recently finalized its Series B funding round.
Niantic Spatial has integrated USDZ export functionality into its Scaniverse application. This addition allows robotics developers to create simulation-ready digital twins from actual physical environments. The tool directly supports use with Nvidia Isaac Sim for advanced simulations. Its primary goal is to mitigate the persistent "sim-to-real" gap in robotics. By bridging this gap, it improves system performance when deploying trained models into complex real-world settings.
Pedestrian Dead Reckoning offers an indoor navigation solution unaffected by GPS signal degradation. The accuracy of foot-mounted AHRS-based PDR heavily relies on the employed attitude estimation algorithm. A Quaternion-Averaging-Based Adaptive Complementary Filter (QAACF) is proposed to improve this accuracy while lowering computational demands. QAACF integrates quaternions from angular velocity with those from acceleration and magnetic field using Markley's quaternion averaging. The method further adaptively adjusts sensor weights based on gait phases and magnetic disturbances, showing lower RMSEs than existing filters with reduced computational overhead compared to Kalman filters.
World Action Models (WAMs) for robotics often fail to capture necessary temporal geometry using appearance-based video latents. MECo-WAM introduces a Multi-Expert Co-Training approach by injecting action-relevant 4D geometric priors into video-action representations. During training, the model uses a lightweight 4D expert supervised by relational targets from a frozen encoder and employs asymmetric visibility to prevent spurious shortcuts. A decayed 4D read-mask attention progressively transfers this geometric knowledge while minimizing deployment cost. Evaluation on datasets like LIBERO and RoboTwin confirms that MECo-WAM improves manipulation performance significantly without increasing inference complexity.
Semantic occupancy memory is crucial for embodied agents navigating connected indoor spaces by structuring representations of occupied regions, free space, unknown areas, and object semantics. Recognizing gaps in current research focusing on single-view or room-level perception, the authors introduce HIOcc, a benchmark unifying multiple indoor datasets into a consistent sparse semantic format supporting local, room-level, and building-level evaluation. They propose GEM-Occ, a Gaussian Evidence Memory framework that treats local visual predictions as transient evidence rather than fixed pointmaps. GEM-Occ converts this evidence into semantic Gaussian occupancy and free-space ray data for persistent storage via causal updates. This hierarchical memory structure—comprising caches, submaps, and a building graph—improves performance across all tested aspects compared to prior baselines on the HIOcc benchmark.
Trajectory planning significantly impacts robot manipulator performance metrics like tracking accuracy and actuator demand. Classical planners are purely kinematic, leading to inefficiencies when executing nominally smooth paths due to ignored system dynamics. This paper introduces a control-aware optimal trajectory planning framework incorporating explicit manipulator dynamics and actuator effort into a finite-horizon formulation. A unique unified evaluation framework tests all planners under identical closed-loop nonlinear execution conditions. Simulations on an UR5 manipulator demonstrate that the proposed method consistently reduces tracking error, corrective torque, and overall execution cost compared to classical approaches.
Accurate vehicle localization is critical for autonomous systems, but performance degrades during sensor outages despite IMU/satellite fusion. This paper introduces Physics-Regularized Machine Learning for Localization (PRML2), a hybrid framework combining Kalman filtering and data-driven learning. PRML2 leverages physics-regularization by training an ML model through a differentiable Kalman filter, ensuring consistency with vehicle motion models. This approach enhances localization accuracy and generalization across various driving conditions. Testing shows PRML2 achieves superior, real-time localization performance using only onboard sensors and introduces a novel dataset for low-friction research.
A new neural network framework proposes transfer learning to model soft, fin-actuated underwater robot dynamics. This approach specializes in handling morphologically similar robots that vary only in scale and hydrodynamic traits. The method utilizes an autoencoder-based domain adaptation technique to create a shared latent representation of the robot dynamics. By aligning this representation, it facilitates efficient knowledge transfer from a larger source robot to a smaller target robot. Testing on two real underwater robots confirmed accurate state estimation on the target platform without requiring labeled data for fine-tuning.
Addressing reliable localization challenges in GPS-denied environments, the authors propose Evidential Velocity Correction using Mamba (EVC-Mamba). This novel, learning-based architecture estimates a virtual velocity sensor directly from onboard vehicle data, eliminating the need for extra hardware. It combines a Mamba selective state space model to capture motion dynamics with evidential deep learning for principled uncertainty quantification. The resultant uncertainty-aware velocity estimate corrects IMU drift when fed into an Error-State Extended Kalman Filter. Testing on real data showed this method achieved localization accuracy within 10% of dedicated external sensors, supporting real-time operation on edge hardware during extended GPS outages.
Multi-agent motion prediction is crucial for safe automated vehicle operation, as previous methods have limitations in mode diversity or prediction accuracy. To overcome these shortcomings, a mode-world weighted regression loss is introduced to improve both world ranking and top-1 confidence while mitigating mode collapse. The proposed framework also incorporates an iterative decoder to enhance prediction accuracy through recurrent and segmental trajectory generation. Empirical testing demonstrates that this novel approach achieves the highest ranking on the Argoverse 2 multi-agent motion forecasting benchmark.
Developed and deployed an autonomous social robot capable of providing Cognitive Stimulation Therapy (CST) at home to address shortages of trained professionals. Nine people living with dementia participated in a one-week study involving daily robot-led sessions. Participants engaged positively, achieving an adherence rate higher than typical rates observed during caregiver-led CST. Family members played a supportive role by initiating sessions and joining activities. This research demonstrates the feasibility and potential for using socially assistive robots to offer scalable in-home cognitive therapy for dementia care.
Calibrating multi-fisheye camera systems is challenging due to increasing complexity and reliance on high-quality observations. Research reveals that calibration failures are primarily caused by intrinsic initialization issues when limited radial span data couples focal scale with fisheye projection parameters, leading to ill-conditioned updates. The paper proposes CO-Calib, a framework integrating a learning-based target detector and an error-analysis-guided frame selector. CO-Calib creates initialization-friendly anchors and multi-camera constraints without altering existing calibration workflows. Experiments show CO-Calib significantly boosts success rates from 68.1% to 99.3% while improving accuracy on various systems.